s_);
  ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);

  const char* tensor_data[] = {&tensor1[0], &tensor2[0]};
  const size_t tensor_size[] = {TF_ARRAYSIZE(tensor1), TF_ARRAYSIZE(tensor2)};
  const int64_t* tensor_dims[] = {&dims1[0], &dims2[0]};
  const size_t tensor_ndims[] = {ndims1, ndims2};
  for (int i = 0; i < 2; ++i) {
    TF_Tensor* v = values[i];

    std::vector<TFE_TensorHandle*> result;
    result.reserve(tensors_.size());
    for (const TensorHandlePtr& tensor : tensors_) {
      result.emplace_back(tensor.get());
    }
    return result;
  }

 private:
  ParallelTensor(const ParallelDevice& device,
                 std::vector<TensorHandlePtr> tensors,
                 absl::Span<const int64_t> shape, const TF_DataType dtype)

      });
}
bool Tape::ShouldRecord(
    absl::Span<const AbstractTensorHandle* const> tensors) const {
  std::vector<int64_t> tensor_ids(tensors.size());
  std::vector<tensorflow::DataType> tensor_dtypes(tensors.size());
  for (int i = 0; i < tensors.size(); i++) {
    tensor_ids[i] = ToId(tensors[i]);
    tensor_dtypes[i] = tensors[i]->DataType();
  }
  return GradientTape::ShouldRecord(tensor_ids, tensor_dtypes);
}

  for (int component_index = 0; component_index < tensors_.size();
       ++component_index) {
    // TODO(allenl): Add a C API for summarizing tensors. Currently custom
    // devices limiting themselves to a C API (for ABI compatibility) would need
    // to implement summarization for component tensors themselves.
    ImmediateExecutionTensorHandle* component =
        tensorflow::unwrap(tensors_[component_index].get());

                       const string& op_name = "");
  // Returns whether any tensor in a list of tensors is being watched and has
  // a trainable dtype.
  bool ShouldRecord(
      absl::Span<const AbstractTensorHandle* const> tensors) const;
  // Unwatches this tensor on the tape. Mainly used for cleanup when deleting
  // eager tensors.
  void DeleteTrace(const AbstractTensorHandle*);

    *   Transpose now supports 6D tensors.
    *   Float LSTM now supports diagonal recurrent tensors: https://arxiv.org/abs/1903.08023

*   `tf.experimental.dtensor`:

    *   Coordination service now works with `dtensor.initialize_accelerator_system`, and enabled by default.
    *   Add `tf.experimental.dtensor.is_dtensor` to check if a tensor is a DTensor instance.

*   `tf.data`:

                                       TFE_TensorHandle* tensor,
                                       TF_Status* status, void* device_info) {
  TF_SetStatus(
      status, TF_UNIMPLEMENTED,
      absl::StrCat("Trying to copy a tensor ",
                   tensorflow::unwrap(tensor)->DebugString(),
                   " on to a parallel device. Pack non-parallel "
                   "tensors for each device into a parallel tensor explicitly.")

    // Input tensors
    const std::vector<std::pair<string, Tensor>>& input_pairs,
    // Output tensors
    const std::vector<string>& output_tensor_names, TF_Tensor** c_outputs,
    // Target nodes
    const std::vector<string>& target_oper_names, TF_Buffer* run_metadata,
    TF_Status* status) {
  const int noutputs = output_tensor_names.size();
  std::vector<Tensor> outputs(noutputs);
  Status result;

// Note: Keep in sync with exported copy of enum in eager/c_api.h.
enum ContextDevicePlacementPolicy {
  // Running operations with input tensors on the wrong device will fail.
  DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  DEVICE_PLACEMENT_WARN = 1,
  // Silently copy the tensor, which has a performance cost since the operation
  // will be blocked till the copy completes. This is the default policy.

  TFE_DeleteContextOptions(opts);

  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  /* Want to test simple MatMul example with abstract tensors:
    [[1,2],   *  [[5,6],   =   [[19,22],
     [3,4]]       [7,8]]        [43,50]]
  */

  // Build 1st Matrix.
  int64_t dims[] = {2, 2};  // Matrices will be 2 x 2